#machine learning

New research from Kaikaku.AI argues that food recommendation systems should distinguish between ingredients that appear together in recipes and those that are chemically similar.

An Interesting Engineering report says a team of computational linguists and archaeologists built a system that can read a 3,500-year-old Hittite script with 90% accuracy.

A new survey argues that World Action Models could help robots move beyond simple image-to-action mapping by simulating how the world is likely to change before they act.

MIT Technology Review has singled out world models as one of the most important areas in AI right now, highlighting rising interest in systems designed to better reason about the real world.

OpenAI’s Parameter Golf challenge drew more than 1,000 participants and over 2,000 submissions, revealing both the creative upside and new governance problems of AI-assisted research.

Google DeepMind has taken a minority stake in the company behind EVE Online and plans to use an offline version of the game to study AI systems that must plan, remember, and learn in complex, persistent environments.

Tutor Intelligence has built a 100-robot ‘Data Factory’ to gather human-guided real-world training data for its vision-language-action model, arguing that scalable robot intelligence requires industrial-grade teaching,

A machine-learning study from VIB and KU Leuven argues that Parkinson’s is not one disease but a family of molecularly distinct disorders, a finding that could reshape how therapies are developed and matched to patients.

A new preprint says a machine-learning analysis of more than 83 million TESS light curves identified 11,554 possible exoplanets, a haul large enough to potentially transform the field if confirmed.

Startup Goodfire says its new Silico system can map internal model pathways and let researchers adjust them during training, an attempt to make large-model development more inspectable and controllable.

An AI News explainer argues that public attention often fixates on outputs while overlooking a deeper question: how models understand information, and how encoders evolved into a core piece of multimodal AI.

MIT Technology Review says DeepSeek’s preview of V4 matters for three reasons: longer prompt handling, performance that challenges closed rivals, and optimization for Huawei Ascend chips.

Google is spotlighting its Tensor Processing Units as demand for larger AI workloads accelerates, framing custom chips as a core advantage in the race to train and run modern models at scale.

A study using electronic health records from more than 10,000 children found machine learning models could identify which patients with early-life atopic dermatitis were most likely to develop persistent asthma or later,

Snowflake is broadening its AI portfolio by expanding Snowflake Intelligence and Cortex Code, aiming to serve both everyday enterprise users and developers building inside the company’s platform.

Researchers are developing a blood test that uses machine learning to analyze circulating DNA linked to liver scarring, with the goal of identifying disease earlier and reducing progression to liver cancer.

A new analysis argues that quantum computers could gain practical advantages in machine learning by loading classical data in smaller batches, sidestepping a long-standing memory bottleneck.

Researchers say machine learning allowed them to track the 2025 Santorini seismic sequence in near real time, revealing more than 60,000 earthquakes and patterns consistent with magma intrusion.

A new robotics boom is being driven by a change in how machines learn, with investors pouring $6.1 billion into humanoid robots in 2025 as the field moves beyond hand-coded behavior

After a breakout year of progress, artificial intelligence is moving from math competitions into frontline research, helping mathematicians generate conjectures, test ideas, and prove new results faster.